DeepLense

This repository contains implementations of ResNet-18, Physics-Informed ResNet-18 (PI-ResNet-18) and PI-ResNet-18 with gradient preprocessing (ResNet-18-Preprocessed) models for classifying dark matter substructures, developed as part of the Machine Learning for Science (ML4Sci) Google Summer of Code (GSoC) 2025 DeepLense evaluation tasks.

• Technical Report

Table of Contents

• Method
• Results
• Notebooks
• Model Weights

Method

Task 1: Standard ResNet-18

• Model: ResNet-18
• Description: A standard ResNet-18 model is implemented to classify dark matter substructures in gravitational lensing images.

Task 2: Physics-Informed ResNet-18

• Model: PI-ResNet-18

• Description: A Physics-Informed ResNet-18 (PI-ResNet-18) integrates gravitational lensing equations into the classification pipeline, utilizing the Physics Block from LensPINN that computes the inverse source image using gravitational lensing equations. The inverse source image is combined with the observed lensing image as a two-channel input to ResNet-18.
  
  

• Model: PI-ResNet-18-Preprocessed

• Description: A gradient map is generated based on lensing images as the third channel of PI-ResNet-18.

Results

The models were evaluated using ROC curves and validation loss metrics across datasets labeled 'no', 'sphere', and 'vort'. Below are the visualizations:

Figure 1: ROC curves showing the classification performance of ResNet-18, PI-ResNet-18, and PI-ResNet-18 with Gradient Preprocessing across the three datasets - no substructure, subhalo substructure, and vortex substructure.

Figure 2: Validation loss over training epochs for the three models, illustrating their convergence behavior.

For details, see the Technical Report.

Notebooks

• task_1_ResNet.ipynb: Implements and trains the standard ResNet-18 model for Task 1.
• task_5_PI_ResNet.ipynb: Implements and trains the PI-ResNet-18 model for Task 5.
• task_5_PI_ResNet_preprocess.ipynb: Implements and trains the PI-ResNet-18-Preprocessed model for Task 5.
• evaluate.ipynb: Evaluates model performance and generates ROC curves for comparison.

Model Weights

Pre-trained model weights are available for download from Google Drive:
Download Model Weights

• ResNet 91.pth: Weights for the standard ResNet-18 model (Task 1).
• PI_ResNet 95.pth: Weights for the PI-ResNet-18 model (Task 5).
• PI_ResNet_Preprocessed 90.pth: Weights for the PI-ResNet-18 model with gradient preprocessing (Task 5).